LSTM-Based Model Compression for CAN Security in Intelligent Vehicles

Abstract

The rapid deployment and low-cost inference of controller area network (CAN) bus anomaly detection models on intelligent vehicles can drive the development of the Green Internet of Vehicles. Anomaly detection on intelligent vehicles often utilizes recurrent neural network models, but computational resources for these models are limited on small platforms. Model compression is essential to ensure CAN bus security with restricted computing resources while improving model computation efficiency. However, the existence of shared cyclic units significantly constrains the compression of recurrent neural networks. In this study, we propose a structured pruning method for long short-term memory (LSTM) based on the contribution values of shared vectors. By analyzing the contribution value of each dimension of shared vectors, the weight matrix of the model is structurally pruned, and the output value of the LSTM layer is supplemented to maintain the information integrity between adjacent network layers. We further propose an approximate matrix multiplication calculation module that runs in the whole process of model calculation and is deployed in parallel with the pruning module. Evaluated on a realistic public CAN bus dataset, our method effectively achieves highly structured pruning, improves model computing efficiency, and maintains performance stability compared to other compression methods.